A to Z of AI: Part 1

A to C

A is for Agentic

Using Agentic AI means allowing AI to do things on your behalf - literally giving it agency. Unlike AI models that create a response in a reaction to a prompt, it can set goals, plan and execute tasks. It can also retain the context of its instruction across longer tasks.

You may have seen posts where people describe booking holidays, doing shopping online or even making trades using Agentic AI. To do so, the user has given Agentic AI access to personal data such as email, social media, calendar, online shopping accounts and payment details.

Maybe you also saw stories about Agentic AI going on an unexpected spending spree, or placing trades beyond the user's budget. Stories like these illustrate the need to carefully think about what it should be permitted to do and what data it should have access to.

Agents can be unpredictable and can be swayed by indirect prompts. By design, they operate best when given access to multiple data sources, so the risk of using data in unintended ways is increased accordingly.

Consider some of the core principles for the responsible use of AI: Fairness, Explainability and Transparency, Accountability, Robustness and Security, and Privacy. Each could be more complex with Agentic AI. How would you explain an autonomous tool's multi-step logic to a regulator? What level of human oversight is there to protect you?

If you're thinking of using Agentic AI in a business context, choose an enterprise-grade platform to maintain a secure, controlled, "human in the loop" environment.

B is for Bias

Bias: when an AI system produces outcomes that systemically favour or disadvantage certain groups.

AI models are just as capable of biased outcomes as humans. This is because AI is trained on data sets that can contain errors or data weighed towards or against one human demographic over others. The training data may also be incomplete.

Around the time of the last Women’s World Cup, there was a lot of media attention on the disproportionate number of professional footballers in the women’s game suffering cruciate ligament injuries compared to the number in the men’s game. Six times as many women suffered cruciate ligament injuries. Some of the articles noted that women’s football boots could be a factor, as it was revealed that the boots had been designed for men, with the only change being the sizing. The boot makers were even asked to appear before a Parliamentary committee to explain.

Back in 2018 Reuters reported that Amazon had scrapped a form of AI it had trialled to sift through job applications. It had been trained on ten years of data from previous applicants. Most of the applicants had been men, so the algorithm started to filter out applications from women.

You might think that removing gender, ethnicity or age from the applications would work, but proxy data points like postcodes, previous job titles and career breaks can still result in the same biased results.

Incomplete data is part of the story. Assuming that the issue is identified before it is replicated at scale, a host of methods of improving data can be applied and the model can be weighted to give greater or lesser importance to particular datasets. It is therefore possible to mitigate bias through data, technical measures and human oversight (“human in the loop”).

In addition to data bias, the AI models can amplify patterns in data, even where the dataset isn’t overtly discriminatory. Equipment and systems that fail to adjust to different skin tones routinely produce worse outcomes in health, policing and security. This raises a serious ethical issue: if a system consistently performs worse for one demographic, can any margin of error be considered acceptable?

Every Responsible AI framework refers to the need for fairness in the use of AI, the creation of outcomes from the use of AI and the use of those outcomes. In the UK there are protections in a broad range of laws, such as equality laws and privacy laws. The EU AI Act (which applies to developers and companies using AI if a person or entity resident in the EU is the subject of AI decision making) specifically calls out fairness.

C is for Compute

AI systems depend on compute, but most of us had barely heard this word until governments started to fret over whether the country will have enough of it.

What compute does: it is the processing power that allows models to analyse data, recognise patterns and generate outputs, e.g. answer your prompts.

What compute comprises: it includes the general‑purpose CPUs (central processing units) found in most computers, and the processors that AI relies on called GPUs (graphic processing units). GPUs were originally designed for graphics but their ability to handle thousands of operations in parallel makes them suitable for AI.

The bigger and more complex AI models become, the more compute is required. Training a frontier‑scale model can require vast amounts of processing power, energy and cooling, but even running smaller models at scale (for search, chat, document review or analytics) depends on access to reliable, efficient compute.

This is also why there are so many stories about data centres in the news. GPUs need a home with substantial fibre connectivity, diverse power feeds, security and cooling equipment. Twenty five years ago a data centre might have had a requirement for a few megawatts of power, but we are now hearing of more and more gigawatt datacentres. More AI goes hand in hand with more compute, more data centres, more power generation and more use of precious resources.

There are clearly practical and ethical questions around an ever increasing demand for compute. Aside from being energy‑intensive and having a colossal environmental impact, it is unevenly distributed. Politics plays a huge part in which countries receive supplies of GPUs, but so does market domination: the biggest AI companies have access to large amounts of compute can build and deploy more powerful systems than the rest of us.

Compute is increasingly seen as a strategic resource. We are already seeing governments restrict access to competing economies. There are legitimate concerns about misuse of compute and whether access to high‑end compute should be regulated.

For most businesses, compute is something we purchase indirectly. This might be through cloud services or because it is embedded in tools we already use.